NVIDIA Publishes GR00T Whole-Body Control Codebase, Including SONIC Models for Unitree G1

NVIDIA’s GitHub repository, “GR00T Whole-Body Control,” includes code, model checkpoints and scripts for training, evaluating and deploying whole-body control for humanoid robots. It contains Decoupled WBC, used in GR00T N1.5 and N1.6, as well as the GEAR-SONIC series and MotionBricks. Three SONIC checkpoints are available for the Unitree G1.

Models for the Unitree G1 offer reference look-ahead horizons of about 200 milliseconds and about 80 milliseconds, and the repository provides C++ and TensorRT deployment options, allowing developers to choose a control configuration for their needs. The look-ahead horizon does not represent total latency, and the repository’s release alone does not confirm performance on a physical robot; both points are important when evaluating potential applications.

The focus of this release is not just an individual robot motion model, but a single repository that covers training, evaluation, data collection through teleoperation and deployment on physical robots. It includes Decoupled WBC, used in GR00T N1.5 and N1.6, alongside the GEAR-SONIC series and SONIC, bringing existing decoupled control and controllers for whole-body motion into one development environment. However, their inclusion does not demonstrate performance in real-world environments or widespread adoption. The three SONIC models are not distinguished solely by speed. The Default version is for general motion tracking, planning and teleoperation; the Low-latency version is for more responsive whole-body teleoperation and VLA execution; and v1.1 uses extended training for robot-heading-based poses and wrist poses. It is also important to note that the approximately 200-millisecond and approximately 80-millisecond look-ahead horizons describe the range of references provided to the controller, not total latency. Confusing them with measured response times could lead to an inaccurate assessment of robot operation. While a report by this newspaper on a real-world demo covered by Robosta with Fujisoft and Daido concerned a physical-robot demonstration, this release shares a development framework for control models and implementation methods; the two provide different kinds of evidence. SONIC brings together models for the G1, deployment using C++ and TensorRT, and data collection for teleoperation. This appears to give developers a way to follow the process of using motion data to train a model and applying its output to a robot through a controller. However, the presence of code and weights in a repository alone does not establish success rates for specific tasks or sustained operation. From a business and supply-chain perspective, the question is how the ability to test multiple control methods and models on a common deployment framework will affect implementation burdens for robot makers and researchers. However, the SONIC models are specifically identified as being for the Unitree G1, and compatibility or performance on other robots is not stated. The model weights are covered by the NVIDIA Open Model License, but the terms for the code and individual components may differ from those for the weights, so users need to check them separately. The next points to verify are reproducibility on different robots, measured performance by task, safety, and the terms for collecting and using training data.

All three SONIC models output 64-dimensional latent motion tokens and have a 50Hz control rate. Default SONIC and SONIC v1.1 use references covering 10 frames at 20-millisecond intervals, for a look-ahead horizon of about 200 milliseconds. Low-latency teleoperation uses four frames, for a look-ahead horizon of about 80 milliseconds, and is intended for whole-body teleoperation and VLA execution. This look-ahead horizon does not represent the latency of the entire teleoperation process, including sensors, communications, preprocessing and inference. Deployment uses C++ and TensorRT, while PyTorch checkpoints support evaluation and continued training in Isaac Lab. The repository also includes tools for collecting data through teleoperation. SONIC model weights are covered by the NVIDIA Open Model License.